Data centers play an important role in today's IT infrastructure. However, their enormous power consumption makes them very expensive to operate. Sadly, much of the power used by data centers is wasted because of poor capacity management, leading to low server utilization.

In order to reduce data center power consumption, researchers have proposed several dynamic server provisioning approaches. However, there are many challenges that hinder the successful deployment of dynamic server provisioning, including: (i) unpredictability in workload demand, (ii) switching costs when setting up new servers, and (iii) unavailability of data when provisioning stateful servers. Most of the existing research in dynamic server provisioning has ignored, or carefully sidestepped, these important challenges at the expense of reduced benefits. In order to realize the full potential of dynamic server provisioning, we must overcome these associated challenges.

This thesis provides new research contributions that explicitly address the open challenges in dynamic server provisioning. We fist develop novel performance modeling tools to estimate the effect of these challenges on response time and power. In doing so, we also address several long-standing open questions in queueing theory, such as the analysis of multi-server systems with switching costs. We then present practical dynamic provisioning solutions for multi-tier data centers, including novel solutions that allow scaling the stateful caching tier and solutions that are robust to load spikes. Our implementation results using realistic workloads and request traces on a 38-server multi-tier testbed demonstrate that dynamic server provisioning can successfully meet typical response time guarantees while significantly lowering power consumption.

While this thesis focuses on server provisioning for reducing power in data centers, the ideas presented herein can also be applied to: (i) private clouds, where unneeded servers can be repurposed for "valley-filling" via batch jobs, to increase server utilization, (ii) community clouds, where unneeded servers can be given away to other groups, to increase the total throughput, and (iii) public clouds, where unneeded virtual machines can be released back to the cloud, to reduce rental costs.

196 pages

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu